Merge branch 'dev_gzf_exp' of github.com:alibaba-damo-academy/FunASR into d...

游雁

2024-04-29 2779602177ae5374547c7a7e17de0b11a166326d

 funasr/train_utils/trainer.py

@@ -15,6 +15,11 @@
from funasr.train_utils.average_nbest_models import average_checkpoints
from torch.distributed.fsdp.sharded_grad_scaler import ShardedGradScaler

try:
    import wandb
except:
    wandb = None


@contextmanager
def maybe_autocast(enabled):
@@ -107,9 +112,23 @@
        self.best_step_or_epoch = ""
        self.val_acc_step_or_eoch = {}
        self.val_loss_step_or_eoch = {}
        
        self.reset_gpu_cache = kwargs.get("reset_gpu_cache", False)

        self.reset_gpu_cache = kwargs.get("reset_gpu_cache", False)
        self.start_data_split_i = 0
        self.start_step = 0
        self.step_in_epoch = 0
        self.use_wandb = kwargs.get("use_wandb", False)
        if self.use_wandb:
            wandb.login(key=kwargs.get("wandb_token"))
            wandb.init(
                config=kwargs,
                project=kwargs.get("wandb_project", "my_project"),
                entity=kwargs.get("wandb_team", "my_team"),
                name=kwargs.get("wandb_exp_name", "my_exp"),
                dir=output_dir,
                job_type="training",
                reinit=True,
            )

    def save_checkpoint(
        self,
@@ -119,6 +138,8 @@
        optim=None,
        scheduler=None,
        scaler=None,
        step_in_epoch=None,
        **kwargs,
    ):
        """
        Saves a checkpoint containing the model's state, the optimizer's state,
@@ -129,6 +150,7 @@
            epoch (int): The epoch number at which the checkpoint is being saved.
        """

        step_in_epoch = None if step is None else step_in_epoch
        if self.rank == 0:
            logging.info(f"Save checkpoint: {epoch}, rank: {self.local_rank}\n")
            # self.step_or_epoch += 1
@@ -142,7 +164,13 @@
                "val_loss_step_or_eoch": self.val_loss_step_or_eoch,
                "best_step_or_epoch": self.best_step_or_epoch,
                "avg_keep_nbest_models_type": self.avg_keep_nbest_models_type,
                "step": step,
                "step_in_epoch": step_in_epoch,
                "data_split_i": kwargs.get("data_split_i", 0),
                "data_split_num": kwargs.get("data_split_num", 1),
                "batch_total": self.batch_total,
            }
            step = step_in_epoch
            if hasattr(model, "module"):
                state["state_dict"] = model.module.state_dict()

@@ -176,7 +204,7 @@
                    )
                else:
                    logging.info(
                        f"No improvement in acc: {self.val_acc_step_or_eoch[ckpt_name]:.4f} < {self.val_acc_step_or_eoch[self.best_step_or_epoch]:.4f}"
                        f"No improvement in acc: {self.val_acc_step_or_eoch[ckpt_name]:.4f} < {self.val_acc_step_or_eoch[self.best_step_or_epoch]:.4f}, {os.path.join(self.output_dir, self.best_step_or_epoch)}"
                    )
            elif self.avg_keep_nbest_models_type == "loss":
                if (
@@ -191,7 +219,7 @@
                    )
                else:
                    logging.info(
                        f"No improvement in loss: {self.val_loss_step_or_eoch[ckpt_name]:.4f} > {self.val_loss_step_or_eoch[self.best_step_or_epoch]:.4f}"
                        f"No improvement in loss: {self.val_loss_step_or_eoch[ckpt_name]:.4f} > {self.val_loss_step_or_eoch[self.best_step_or_epoch]:.4f}, {os.path.join(self.output_dir, self.best_step_or_epoch)}"
                    )
            else:
                print("Undo")
@@ -232,7 +260,7 @@
            ckpt = os.path.join(self.output_dir, "model.pt")
            if os.path.isfile(ckpt):
                checkpoint = torch.load(ckpt, map_location="cpu")
                self.start_epoch = checkpoint["epoch"] + 1
                self.start_epoch = checkpoint["epoch"]
                # self.model.load_state_dict(checkpoint['state_dict'])
                src_state = checkpoint["state_dict"]
                dst_state = model.state_dict()
@@ -268,6 +296,17 @@
                self.best_step_or_epoch = (
                    checkpoint["best_step_or_epoch"] if "best_step_or_epoch" in checkpoint else ""
                )
                self.start_data_split_i = (
                    checkpoint["data_split_i"] if "data_split_i" in checkpoint else 0
                )
                self.batch_total = checkpoint["batch_total"] if "batch_total" in checkpoint else 0
                self.start_step = checkpoint["step"] if "step" in checkpoint else 0
                self.start_step = 0 if self.start_step is None else self.start_step
                self.step_in_epoch = (
                    checkpoint["step_in_epoch"] if "step_in_epoch" in checkpoint else 0
                )
                self.step_in_epoch = 0 if self.step_in_epoch is None else self.step_in_epoch

                model.to(self.device)
                print(f"Checkpoint loaded successfully from '{ckpt}'")
            else:
@@ -295,7 +334,7 @@
        """
        if self.use_ddp or self.use_fsdp:
            dist.barrier()
        logging.info(f"Train epoch: {epoch}, rank: {self.local_rank}\n")
        logging.info(f"Train epoch: {epoch}, rank: {self.rank}\n")
        model.train()

        # Set the number of steps for gradient accumulation
@@ -315,6 +354,7 @@
                if iterator_stop > 0:
                    break
            self.batch_total += 1
            self.step_in_epoch += 1
            time1 = time.perf_counter()
            speed_stats["data_load"] = f"{time1-time_beg:0.3f}"

@@ -327,7 +367,7 @@
                time2 = time.perf_counter()
                with maybe_autocast(self.use_fp16):
                    retval = model(**batch)
                    

                    if (
                        self.reset_gpu_cache
                        and (torch.cuda.max_memory_reserved() / 1024 / 1024 / 1024) > 70
@@ -417,6 +457,7 @@
                self.log(
                    epoch,
                    batch_idx,
                    step_in_epoch=self.step_in_epoch,
                    batch_num_epoch=batch_num_epoch,
                    lr=lr,
                    loss=loss.detach().cpu().item(),
@@ -428,16 +469,17 @@
                    data_split_num=kwargs.get("data_split_num", 1),
                )

            if (batch_idx + 1) % self.validate_interval == 0:
            if self.step_in_epoch % self.validate_interval == 0:
                self.validate_epoch(
                    model=model,
                    dataloader_val=dataloader_val,
                    epoch=epoch,
                    writer=writer,
                    step=batch_idx + 1,
                    step_in_epoch=self.step_in_epoch,
                )

            if (batch_idx + 1) % self.save_checkpoint_interval == 0:
            if self.step_in_epoch % self.save_checkpoint_interval == 0:
                self.save_checkpoint(
                    epoch,
                    model=model,
@@ -445,6 +487,9 @@
                    scheduler=scheduler,
                    scaler=scaler,
                    step=batch_idx + 1,
                    step_in_epoch=self.step_in_epoch,
                    data_split_i=kwargs.get("data_split_i", 0),
                    data_split_num=kwargs.get("data_split_num", 1),
                )

            time_beg = time.perf_counter()
@@ -474,7 +519,7 @@
        """
        if self.use_ddp or self.use_fsdp:
            dist.barrier()
        logging.info(f"Validate epoch: {epoch}, rank: {self.local_rank}\n")
        logging.info(f"Validate epoch: {epoch}, rank: {self.rank}\n")
        model.eval()

        with torch.no_grad():
@@ -552,10 +597,10 @@
                    iterator_stop.fill_(1)
                    dist.all_reduce(iterator_stop, dist.ReduceOp.SUM)

        if kwargs.get("step", None) is None:
        if kwargs.get("step_in_epoch", None) is None:
            ckpt_name = f"model.pt.ep{epoch}"
        else:
            ckpt_name = f'model.pt.ep{epoch}.{kwargs.get("step")}'
            ckpt_name = f'model.pt.ep{epoch}.{kwargs.get("step_in_epoch")}'
        self.val_acc_step_or_eoch[ckpt_name] = self.val_acc_avg
        self.val_loss_step_or_eoch[ckpt_name] = self.val_loss_avg
        model.train()
@@ -568,6 +613,7 @@
        self,
        epoch=0,
        batch_idx=0,
        step_in_epoch=0,
        batch_num_epoch=-1,
        lr=0.0,
        loss=0.0,
@@ -598,10 +644,10 @@
            acc_avg_epoch = getattr(self, f"{tag}_acc_avg")
            description = (
                f"{tag}, "
                f"rank: {self.local_rank}, "
                f"rank: {self.rank}, "
                f"epoch: {epoch}/{self.max_epoch}, "
                f"data_slice: {data_split_i}/{data_split_num}, "
                f"step: {batch_idx + 1}/{batch_num_epoch}, total step: {self.batch_total}, "
                f"step_in_slice: {batch_idx + 1}/{batch_num_epoch}, step_in_epoch: {step_in_epoch}, total step: {self.batch_total}, "
                f"(loss_avg_rank: {loss:.3f}), "
                f"(loss_avg_epoch: {loss_avg_epoch:.3f}), "
                f"(ppl_avg_epoch: {math.exp(loss_avg_epoch):.3e}), "
@@ -613,18 +659,29 @@
            )
            logging.info(description)

            description_dict = {
                f"rank{self.rank}_loss/{tag}": loss,
                f"rank{self.rank}_lr/{tag}": lr,
            }

            if writer is not None:
                writer.add_scalar(f"rank{self.local_rank}_loss/{tag}", loss, self.batch_total)
                writer.add_scalar(f"rank{self.local_rank}_lr/{tag}", lr, self.batch_total)
                writer.add_scalar(f"rank{self.local_rank}_lr/{tag}", lr, self.batch_total)
                writer.add_scalar(f"rank{self.rank}_loss/{tag}", loss, self.batch_total)
                writer.add_scalar(f"rank{self.rank}_lr/{tag}", lr, self.batch_total)
                for key, var in stats.items():
                    writer.add_scalar(
                        f"stats_rank{self.local_rank}_{key}/{tag}", var.item(), self.batch_total
                        f"stats_rank{self.rank}_{key}/{tag}", var.item(), self.batch_total
                    )
                    description_dict[f"stats_rank{self.rank}_{key}/{tag}"] = var.item()
                for key, var in speed_stats.items():
                    writer.add_scalar(
                        f"stats_rank{self.local_rank}_{key}/{tag}", eval(var), self.batch_total
                        f"stats_rank{self.rank}_{key}/{tag}", eval(var), self.batch_total
                    )
                    description_dict[f"stats_rank{self.rank}_{key}/{tag}"] = eval(var)
            if self.use_wandb and wandb is not None:
                wandb.log(
                    description_dict,
                    setp=self.batch_total,
                )

    def close(self, writer=None):